RL on coding tasks transfers across benchmarks
The point
Researchers post-trained Kimi K2.7 Code with reinforcement learning alone on 1,700 expert-built coding tasks. Pass rates rose on all six external benchmarks evaluated, across three agent harnesses.
Why it matters
For builders, RL on well graded tasks can teach skills that transfer beyond the training set.
Key details
- Training: 1,000 repository tasks plus 700 terminal tasks, reward drops to zero if any pass-to-pass test fails
- Gains: SWE-Bench Pro 60.1 to 64.8, DeepSWE 31.0 to 43.4, Terminal-Bench 2.1 67.4 to 82.0
Original ~10 min → 25 sec here Source: arXiv cs.AI, Nov 16, 2026